Canva Craft Image Model: What It Does Today and What Is Still Coming

On October 22, Canva introduced Canva Craft, which it calls its first image generation model and the first foundation model its in-house lab, Canva Research, has built and trained from scratch. The Canva Craft image model works differently from a typical generator: Canva says it first writes a plan of the composition and only then renders the picture. It is in Canva's Dream Lab today for all users, but several pieces of the launch are announced rather than live, and the announcement gives no price per image. Here is what you can use now, what is still coming, and when it makes sense to try Craft first.

What Craft does differently

Conventional image models, in Canva's words, "generate pixels directly from a prompt." Craft splits the job in two. According to the announcement, a planning model first reads your prompt and writes a detailed plan covering which elements appear, where they sit relative to each other, how big they are, the visual hierarchy, and any text that has to appear in the image. A separate rendering model then turns that plan into the final image at up to 2K resolution.

Canva's stated payoff is layouts that hold together: the company says modelling composition and typography before rendering lets Craft produce complex layouts with "legible text, precise object placement and consistent visual relationships." It also says the structure allows targeted changes to one element without redoing everything else. These are Canva's claims. I have not tested the model; the one outside hands-on comparison I have read, a three-prompt test by Business Insider, is covered with the rankings below.

The training data explains the focus. Canva says it includes millions of human-made templates from its content library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts. After pre-training, Canva says it refined Craft with supervised fine-tuning and reinforcement learning, using feedback from its professional designers on instruction following, typography and visual quality.

A note on "first." Both of Canva's "firsts" are scoped: the first image generation model, and the first foundation model Canva Research built from scratch. Craft is not Canva's first AI model of any kind: in October 2025 the company launched a design model that, as TechCrunch reported at the time, "generates designs with editable layers and objects rather than flat images." Canva also acquired Leonardo.ai in 2024, per TechCrunch.

How to use Canva Craft today

Canva says Craft is available to all Canva users globally from launch day, and the announcement names no plan restriction. To reach it:

1. Open Dream Lab in Canva. Canva's announcement points you to the bottom left panel, and Canva's Dream Lab help page gives the route from the Homepage: select More, then Dream Lab. It is not inside the editor. TechCrunch also describes access through the Dream Lab tab.
2. Write your prompt. Canva says prompts can be in any language.
3. Pick one of more than 30 aspect ratios, or enter your own, for anything from a story to a banner.
4. To edit the result piece by piece, select the generated image and use Magic Layers, the existing Canva feature that Canva says can turn generated images into editable design elements. Canva says Craft's outputs are "particularly well suited" to it because of the planning structure.

On text inside images, Canva says Craft can create text in 55 languages. The announcement does not list those languages, so I would not assume any specific one, Hebrew included, until you see it render correctly.

Live now versus only announced

The launch mixes current features with future ones, and one sentence in Canva's own announcement blurs the line. This is the status according to Canva's announcement and API documentation:

Feature	Status	Source detail
Craft in Dream Lab	Live for all users	Announcement, October 22
Up to 2K output, 30+ aspect ratios or custom	Live in the Canva product	Announcement
Public API (craft_v1_2k)	Live as a preview	Canva's API reference lists the model and labels the API a preview
Canva MCP (ChatGPT, Gemini and other assistants)	Coming soon	Section headed "Canva MCP (Coming Soon)"; TechCrunch says Canva "plans to support" ChatGPT and Gemini
Craft-fast and Craft-ultrafast	Upcoming	Canva says they are designed for high quality "at a fraction of the cost of leading frontier models"
Craft reference-image capabilities and more advanced editing	Future releases	Announcement. Dream Lab's help page already documents adding a reference image, without naming a model

The MCP row needs an explanation. MCP (Model Context Protocol) is a standard that lets AI assistants call outside tools. Canva says that through Canva MCP, agents will be able to create images that are ready to use in a design, and TechCrunch reports the plan covers third-party chatbots such as ChatGPT and Gemini. One sentence in Canva's announcement says Craft "is also available via Public API and the Canva MCP Server," but the section that describes MCP is explicitly labeled "Coming Soon," and TechCrunch reports it as a plan. I am treating MCP access as not yet live.

For developers: the API is narrower than the editor

Canva's image generation API reference lists craft_v1_2k as a model value, alongside lucid_origin and z_image_turbo. A few details matter before you build on it:

• Six aspect ratios, not 30+. The API accepts square (1:1, the default), landscape (16:9), portrait (9:16), 3:2, 2:3 and 4:5. The 30+ presets and custom ratios are a product feature.
• Preview status. Canva says the API "is currently provided as a preview" and warns of possible unannounced breaking changes.
• Model selection is optional. Canva recommends omitting the model parameter so it picks the best available model and falls back if one is unavailable. If you specifically want Craft, request craft_v1_2k.
• Limits and cost. Requests are rate limited to 20 per minute for each user of your app, and each job consumes the user's AI credit allowance. When that runs out, the API returns a 429 error with the credit_quota_exceeded code.
• Output size varies by model. The job result reports the actual width and height.

What Canva's rankings say, and what they do not

Canva says Craft debuted as the fourth most advanced image generation model on DesignArena, which it describes as "the largest crowdsourced assessor of AI design." Two numbers are easy to mix up here. Fourth is the model's overall rank. Separately, Canva says the result makes Canva Research the second most advanced lab for image generation. In subcategories, Canva reports Craft at #1 for People & Portraits (out of 87 models) and Landscapes & Nature (out of 85), #2 for Abstract & Patterns, and #4 for Products.

These ranks are Canva's report. The DesignArena image leaderboard showed no model rows, scores or date when checked on October 23, so I could not confirm the figures or see which three models rank above Craft overall. Canva also says its designers and researchers compared more than 25,000 image pairs against leading models internally, but the announcement gives no numeric results from that evaluation.

On October 24, Business Insider's Katherine Li compared Craft with models from OpenAI, Google and Meta on three prompts of rising difficulty. In her test, Craft's birthday card had weak typography and an oversized character head, its pendant edit placed the jewelry implausibly, and its photorealistic cat was well composed but had unnatural-looking fur. One small comparison is not a general ranking, and it says nothing about Hebrew text quality.

For these reasons I am not saying Craft beats or trails any specific OpenAI or Google model. If you want to look at other generators, my AI image generator comparison and Nano Banana 2 review cover those models on their own terms.

Canva is also open about weak spots. It says it will keep raising the bar on quality, detail and control to improve "things like fine print and technical drawings."

What is still unknown

• Price and credits. The announcement gives no price, credit cost per image or limit per plan. The Dream Lab help page says generating images, reference images and Boost resolution consume your AI allowance, but its allowance notes give no per-plan numbers. The API docs confirm that jobs there draw on the user's AI credit allowance.
• Independent quality results. Beyond Business Insider's three prompts, Canva's DesignArena ranks were not readable on the leaderboard when I checked, and its internal 25,000-pair evaluation has no published numbers.
• Which languages render. Canva gives a count of 55 but no list.
• Release dates for MCP access, the faster variants and Craft's reference-image capabilities.

Like its other AI products, Canva says Craft is backed by Canva Shield, its approach to AI safety.

When to try Craft first

My rule is based on workflow fit, because the quality evidence I have is Canva's own report of its DesignArena results, which I could not confirm on the leaderboard, plus Business Insider's three-prompt comparison. I recommend reaching for Craft first when all three of these are true:

1. You are already building the design in Canva, so the image lands where you will use it.
2. The image is a layout rather than a single picture: a poster, social post, menu or slide with headline text, which is the kind of material Canva says it trained on.
3. You want to adjust parts afterwards through Magic Layers, or you need a size that the 30+ presets or a custom ratio cover.

Stay with a generator you have already tested for the job when you need fine print or technical drawings (areas Canva itself names for improvement), text in a language you have not seen Craft render, reference images with Craft (announced for future releases), or predictable per-image costs for a large batch (not published). Developers should build only against the six API aspect ratios and plan for preview-stage changes.

The simplest next step is a single test: open Dream Lab, ask for a poster with a short headline in your language at a size you actually use, check for a model label that confirms Craft generated it, and see whether Magic Layers separates the text cleanly. That one result will tell you more about your own workflow than the leaderboard claims can.